Skip to content

RisQ

Summary

RisQ (Risk Query) is a transformer-based multimodal framework designed to learn a unified representation of participant health across multiple diseases, modalities, and time. By integrating structured clinical features, medications, and longitudinal diagnostic histories, RisQ uses a query mechanism with natural language and continuous time representations to predict risk for arbitrary diseases and time horizons. The framework demonstrated state-of-the-art predictive performance when trained on the UK Biobank and validated on the independent All of Us cohort (as of 2026-07-09).

Architecture and Methods

RisQ employs a multi-component architecture to learn a unified representation of participant health: - Input Tokenization: Structured inputs (demographics, lifestyle, environment, physical measurements, blood & urine assays) are processed using feature-specific tokenizers. Diagnoses and medications are encoded with large language models and augmented with continuous sinusoidal encodings representing the time since diagnosis. - Transformer Encoder: Learns a compact health representation by aggregating tokenized inputs using self-attention. - Disease-Risk Decoder: A cross-attention decoder that estimates disease-specific risk. It takes both the patient's health tokens and a query token as input. Each query token combines an LLM embedding of the target disease with a temporal token representing the desired prediction horizon. - Training Objective: The model optimizes a horizon-conditional training objective via binary cross-entropy, learning event probabilities relative to sampled prediction horizons rather than standard time-to-event regression.

Applications and Capabilities

  • Cross-Disease & Zero-Shot Prediction: RisQ achieves strong zero-shot performance on unseen diseases by querying its natural-language target interface, outperforming demographic baselines.
  • Extrapolation Over Time: The framework can extrapolate risk prediction to time horizons beyond those observed during training.
  • Rare Disease Risk Modeling: By jointly modeling multiple diseases, RisQ facilitates information transfer, resulting in substantial prediction gains for low-prevalence diseases (<0.1% event rate) compared to single-disease models.
  • Genetic Burden Association: Exome-wide gene-level loss-of-function (LoF) burden analysis linked rare genetic variants in genes such as OMA1, HBB, LDLR, CASR, and SLC22A12 to cross-disease risk profiles.

Contributing Research Groups

  • Casale Lab — statistical-genetics and machine-learning methods for complex phenotypes.
  • Rueckert Lab — AI methods for healthcare and cohort-scale clinical data.
  • Theis Lab — computational biology and multimodal representation learning.

Citations

  • Hager, P., Roth, B., Bühler, N., Lu, D., Burks, J. H., Shilova, L., Kfuri-Rubens, R., Roellin, E., Pan, J., Di Folco, M., Chan, E., Schnabel, J. A., Adams, L., Rueckert, D., Theis, F. J., & Casale, F. P. (2026). Learning the shared structure of human health across diseases, modalities, and time. medRxiv preprint. DOI: 10.64898/2026.07.07.26357373. Source paper: 2026.07.07.26357373v1.full.pdf